Over-subscription Planning with Numeric Goals J. Benton Computer Sci. & Eng. Dept. Arizona State University Tempe, AZ Minh Do Palo Alto Research Center (PARC) Palo Alto, CA Subbarao Kambhampati Computer Sci. & Eng. Dept. Arizona State University Tempe, AZ
Over-subscription Planning Goals optional & have utility Actions have cost Maximize utility-cost “Benefit” cost = 200 cost = 500 cost = 300 Util = 500 Util = 200 B C A Initial: At A Goals: B & C [“The Mystery Talk”, Smith 2003] Rovers Example 300
Motivation Numeric goals also have utility More soil gives better instrument reading More packages give more profit Cost for achieving varying values differs More soil requires more weight More packages require more deliveries
Objective Want more/less G = soil-sample ∈ [2,4] U(G) = (* (soil-sample) 2) Challenge – A measurable level of numeric goal achievement: degree of satisfaction Collect Cost=1 Collect Cost=2 1 gram cost=3 soil collected util=2*2=4 Collect Cost=3 1 gram action cost cost=6 util=3*2=6 Benefit=4- 3=1 Benefit=6-6=0 Satisfy numeric goals at different values to give varying utility BenefitBenefit v a l u e best benefit
Modeling Numeric Goal Over-subscription Achieve with a given utility Specify a goal range U(G) = (* (soil-sample) 2) G = soil-sample ∈ [2,4] Sample UtilityUtility 1. Fixed utility for satisfying level 2. Linear 3. Hard bounds Infinity on range OK 4. Model as a separate goal
Sapa Mps Architecture Over-subscribed Planning Planning Problem Input Initial State Select state with best f-value Queue of Time-Stamped States Better benefit plan? Yes Output Plan Generate States by Applying Actions Build RTPG Propagate Cost Find Utility No Anytime A* Search Based on Sapa PS
Challenge – Heuristic Support Heuristic needs to… Estimate cost of achieving variable values Find the utility of the values Extend current state-of-the-art techniques Planning graph structure Reachability estimation Cost propagation
Challenge – Find Goal Achievement Cost Propagate reachable values with cost Sample_Soil Communicate Move(Waypoint1) Sample_Soil cost( ): Cost of achieving each value bound v 1 : [0,0] [0,1] [0,2] A range of possible values
Cost Propagation on Variable Bounds Bound cost dependent upon action cost previous bound cost - current bound cost adds to the next Cost of all bounds in expressions Sample_Soil Cost(v 1 =2) Sample_Soil C(Sample_Soil)+Cost(v 1 =1) v 1 : [0,0] [0,1] [0,2] Sample_Soil Cost(v 1 =6) Sample_Soil C(Sample_Soil)+Cost(v 2 =3)+Cost(v 1 =3) v 1 : [0,0] [0,3] [0,6] v 2 : [0,3] Sample_Soil Effect: v 1 +=1 Sample_Soil Effect: v 1 +=v 2
Extracting Relaxed Plan with Numeric Info Start with best benefit bounds Relaxed plan includes Actions Supporting bounds BenefitBenefit v a l u e best benefit
Sample_Soil 1 (Sa1) Dur = 1 Cost: 1 (at end) V 1 += 1 Sample_Soil 2 (Sa2) Dur = 1.25 Cost: 2 (at end) V 1 += 2 Communicate (Com) Dur = 1.5 Cost: 3 (at start) V 1 ≥ 1 Sa1 t C:1 Sa1 C:1 Sa1 C:1 Sa2 C:2 Sa2 C:2 Sa2 C:2 Com C:4 Com C:4 4 Goal: v2 ∈ [5,∞], U(v2 ∈ [5,∞]) = v2 * 3 (at start) V 2 := V 1 v1v1 value cost value cost v2v2 upper time point v 1 – soil sample in rover’s store v 2 – soil sample communicated
Sample_Soil 1 (Sa1) Dur = 1 Cost: 1 (at end) V 1 += 1 Sample_Soil 2 (Sa2) Dur = 1.25 Cost: 2 (at end) V 1 += 2 Communicate (Com) Dur = 1.5 Cost: 3 (at start) V 2 := V 1 (at start) V 1 ≥ 1 Sa1 t C:1 Sa1 C:1 Sa1 C:1 Sa2 C:2 Sa2 C:2 Sa2 C:2 Com C:4 4 v1v1 value cost value cost Com C:4 satisfies goal h(S) = U(G) - (cost of actions + cost of bounds) v2v2
Results – Modified Rovers Added numeric variables: Soil and rock sample amount in rover store More communicated soil/rock - greater utility
Average improvement: 3.06 Results – Modified Rovers
Anytime A* Search Behavior
Results – Modified Logistics Added numeric variables: Number of packages at location More packages - greater utility
Results – Modified Logistics Average improvement: 2.88
Summary Over-subscription planning in the presence of Numeric goals Durative actions Propagating cost over numeric values
Future Work Delayed satisfaction of goals Goal utility dependency late -10 late -10
Questions.